Péter Megyesi

@Yuchenj_UW While are appreciate the OSS model and also believe that in 2-3 years they will be the default, this one I'll give the credit to GPT 5.6 Sol. Currently most of the subscription users would shift to Codex not to OSS alternatives, if the wouldn't let use Fable.

Thank you, Kimi K3 and GLM-5.2!! oss models made it.

Thank god there's competition in the market. My guess is that they didn't want to do this, but had to, otherwise too many people would have switched to Codex.

Wow this score of Kimi K3 on DeepSWE is brutal. If this is an authentic result (meaning they did not trained directly for this bench), then the Moonshot team really cooked something special.

I wonder why @Kimi_Moonshot is not doing distilled versions of Kimi K3. Like with a Kimi-medium 500B and Kimi-small 80B they could rule the whole local AI community if the relationship is similar to let's say Sol - Terra - Luna.

This whole data retention vs. training on the data is so confusing. My current stand on this you need to use the API to be sure, and also it make sense to use them via the cloud provider offerings (Bedrock, AI Foundry, Vertex). Seems they have stronger policies, and they probably aslo cannot give the data to a 3rd party.

Right now it feels like the single biggest competitive advantage an AI lab could have is making it abundantly clear whether and how they will train models on your data I pay pretty close attention to this and I couldn't confidently summarize the policies for ANY of the lead labs

"Early in the rollout of the technology, employees also faced restrictions on using Gemini to write or analyze software over concerns that proprietary code could leak into the AI model’s training data, they said." ... concerns about their OWN code being trained on?

I'm seeing a lot of Kimi K3 post, but where's the official announcement?

Ah yeah, that happend, I believed that was Sonnet 4.5 vs. Opus 4.1. And Google also in this this phase where Flash 3.5 is better then Pro 3.1 (not that anybody uses those model...) But still the question is: let's they make Opus 5.1 right away and it is better and cheaper then Fable 5. Wouldn't they kill their most precious model with this move?

I recall Sonnet 4.5 was superior to Opus what was that 4.3? Anthropic has a history of releasing next gen lower tier model being superior to the old gen higher tier model. I won’t be surprised if Opus 5 exceeds Fable 5 slightly at a much cheaper price benefiting the users with improved efficiency.

Normally I don’t comment on rumors, but if Kimi K3 actually beats out Opus 4.8 that’s nuts Also hearing Opus 5 might drop?

I feel like Anthropic likes timing these reset just a few hours after my own weekly reset 😭

@photonclock @mattpocockuk Yeah, but that is the real trade off: not every task is hard: so compared to Sol max thinking Sol on just high is only 5% worse / less intelligent, but for half the cost and probably 2x speed (finishes faster since generates less token in less agent turns).

@M3gy0 @mattpocockuk This chart just convinced me I need to use Sol Max for everything. 😅

When they're playing with the effort dial, people seem to start high and go low if they need to IMO this is a crazy way to do things - you should start low and only go higher if you need to Effort is essentially throwing more tokens at the problem. This works great in benchmarks, where spending an extra 20% on tokens to yield a 2% improvement is great for marketing. But in regular tasks, such as exploring a repo or modifying a test, it just means burning pointless tokens. This is especially true in Fable-class models where tokens cost so much more. And the more tokens you burn, the faster attention degradation creeps up, and the worse your tokens become. People are a little too convinced by those graphs that show the quality line going up when you increase the effort. So they think "I'll try this out on xhigh, but not max, that seems reasonable" But those graphs hide the cost, the latency, and the unbelievable waste. If you've ever asked the agent to explore your repo and ended up with 200K tokens spent, you know what I mean. So whenever you're trying a new model, try the lowest effort level first. If you see a dip in quality, creep up.

@Jakekrajewski @mattpocockuk You mean Sol, not Sonnet, right? 🤔

@M3gy0 @mattpocockuk The findings from this chart: do everything with Sonnet 5 Max.

Where is Google with the new Gemini model? One year ago everyone thought they might become the winner of the AI race. Today they seem so much behind: xAI and Meta both came out with respectfull models on par(ish) with Opus 4.8 and GPT-5.5, and there's even a chinese open weight model (GLM-5.2) that seems to be better then Gemini. I used Gemini for vision usecases, but today I just switched all of these to Luna.

@mattpocockuk This is a really great benchmark, even Artificial Analysis put this into it's new 4.1 intelligence measurement. Also check out the output token and agent steps tab, very informative.

@M3gy0 Found it, great nudge: deepswe.datacurve.ai

Muse Spark 1.1 got a very good score on DeepSWE. Based on this it is on par for coding with Opus 4.8 and GPT-5.5. This convinced me to dig into this model deeper...

@mattpocockuk Which model are you using, and where's it's dumb zone? Generally I do not like to go above 200k with Opus/Fable

Sometimes I experiment with going deep in the dumb zone to implement a feature Today it cost me 90 minutes of debugging to fix its stupid mistakes in a complex build Need to stop trying this

Yes, I've used this, but I just noticed that there were quite a few updates since I last tried. The current implementation seems good enough, I can see that I can move object (e.g. textboxes), but the layout seems fixed, so it seems I can only reorganize the boxes, I can really do something like move this box 5px up. I guess this is comming from the html div structure that Claude Design uses, and probably the issue is there with all the website build platforms (like WordPress, Wix, etc). But honestly I'm not even sure this is a huge problem, since the layouts seem good enough by default (whereas in the PPT days I always mixed up my templates so I needed these minor adjustments). All in all, I will do a lot more Claude Design when the semester starts in September, so I'll make sure to drop a new feedback if I find somehing fustrating.

@M3gy0 @nateparrott Hey Peter thanks for sharing! Have you tried the "edit" mode in a project - these have some of the direct edit manipulation. Or are you looking for specifically more layout movement related ones?

folks who’ve tried Claude Design — how was the experience? what works well and what could be better?

Just updated Codex and immediately lost browser control. 😑 The bundled Browser/Chrome plugin now crashes on startup with "Cannot redefine property: process" before it can even open a page. At least the agent found that I'm not alone with this: github.com/openai/codex/i…